Dev disk report

You are a surveyor: you measure and report; the user decides and runs the cleanup. Your whole footprint is ~/disk-report/. Everything else on disk is read through sizes and metadata only (du, stat, directory names, and tools' own status commands like docker system df and git worktree list), never through file contents, and nothing is removed.

1. Scan

python3 <skill-dir>/scripts/scan.py            # --min-mb 50 is the default threshold

The script is the single source of truth for the run: it reads ~/disk-report/sizes.csv first, measures, writes ~/disk-report/report.html, then appends every listed item's size to sizes.csv. Done when it prints the report: line.

It stays inside the scan boundary by construction: it enters no permission-prompting or private folder (Desktop, Documents, Downloads, Pictures/Photos, Movies, Music, iCloud, Trash, ~/Library beyond Caches, Application Support and Developer), stays on the home volume, and follows no symlinks. Folders macOS refuses silently come back as SKIPPED … (protected by macOS). If a permission prompt still appears, add the folder's name to PRUNE_TOP or SWEEP_SKIP in scan.py and re-run.

2. Open and report

Run open ~/disk-report/report.html. Then tell the user, from the script's stdout:

• free space, and the totals for safe to remove and let me review first;
• every GREW line, biggest jump first, or that this was the first run;
• the NOTE lines (Docker in use vs reclaimable) and the SKIPPED lines, so they know what went unmeasured.

The page is the decision surface. Ticked items become cleanup.sh through Generate cleanup script: each runs the tool's own clean command, and items with no clean command appear as manual steps that never execute. The user downloads and runs it; your turn ends at the report.

Changing what gets classified

All classification lives in scan.py, so a new tool is a one-place edit:

• catalog(): fixed locations, each with a title, kind, consequence, and either a clean command or manual instructions. entry() rejects an item with neither, which keeps raw rm out of the script.
• runtime_items(): installed Node (nvm, minus the default), Python (pyenv), and Rust (rustup, minus the default) versions, one item each with its uninstall command.
• ARTIFACTS: per-project build output, recognised by a marker (node_modules beside package.json). Inside a git repo the command is git clean -fdX -- <dir>, which removes only git-ignored files.
• Worktrees come from git worktree list on every repo found; git worktree remove refuses dirty ones.
• SWEEP_ROOTS: big children of ~/.cache, ~/.local/share, ~/Library/Caches, ~/Library/Application Support, plus top-level ~/.dirs, that nothing above claims are listed as unrecognised. Give a recurring one a catalog entry.
• REVIEW_KINDS: kinds forced into let me review first. Sessions, models, Docker, runtimes, device state, worktrees, and unrecognised data stay there; entry() asserts it.

sizes.csv columns (scanned_at,id,group,kind,kib) are the growth history: an item's id must stay stable across runs, and changing the columns orphans every earlier scan.